A method is presented for generating statistical models of timing data continuously over very long monitoring sessions. This method is intended for memory-efficient runtime modeling of timing properties in embedded software systems, such as execution times or inter-arrival times, but is a quite generic method that should be applicable for other purposes and domains as well. Specifically, we intend to use this method as a component in automatic generation of simulation models for probabilistic timing analysis of complex embedded software systems. Given a stream of data as input, this method gradually builds up a statistical model capturing the approximate distribution of the data. The method uses a modest and fixed amount of on-target RAM, decided by the desired accuracy of the model, and allows for long monitoring sessions covering billions of data points. The paper presents the motivation, algorithm, a prototype implementation and evaluation using real execution time data from an ARM7 microcontroller.